As a member of the technical team focused on foundation model performance, you will play a key role in understanding, evaluating, and improving the capabilities of frontier foundation models for embodied AI and autonomous scientific experimentation. This spans both autoregressive foundation models — the sequence-modelling backbone behind language, vision-language and action prediction — and world models and world action models that learn the dynamics of the laboratory well enough to simulate, plan and act on experiments before they are run. Your research will systematically investigate how training data, model development stages, and training strategies interact to determine model Internal capability. By uncovering these interactions, you will identify performance bottlenecks and develop novel approaches, such as new reward models, new learning curricula or data mixing strategies, that continuously improve foundation model performance.
This role offers a unique opportunity to conduct frontier research at the intersection of large-scale foundation models, data-centric AI, and embodied intelligence while solving real-world scientific problems. You will work closely with other AI researchers, software engineers, robotics engineers, and domain scientists to translate advances in autoregressive foundation models and world models into measurable improvements in autonomous laboratory performance.
The successful candidate will design and execute systematic experimental studies to understand how different data mixtures, data quality, model architectures, and training stages influence downstream capabilities and final performance in an end-to-end scientific workflow. You will identify performance bottlenecks and conduct research on improving foundation model performance through a deeper understanding of the interactions between data and models.